P2P technologies, PlanetLab, and their relevance to Grid work

Matei RipeanuThe University of Chicago

1

10

100

1000

10000

1 10 100 1000Rank (log scale)

LinP

ack

GFL

OPS

(log

scal

e) .

2001

Why don’t we build a huge supercomputer?

1

10

100

1000

10000

1 10 100 1000Rank (log scale)

LinP

ack

perf

.GFL

OPS

(log

scal

e) . 2001 2000

1999 19981997 19961995

Top500 supercomputer list over time:Zipf distribution: Perf(rank) ≈ rank -k

Parameter 'k' evolution .

-0.84-0.82-0.80-0.78-0.76-0.74-0.72-0.70-0.68

1995

1996

1997

1998

1999

2000

2001

2002

2003

Trend

Increasingly interesting to aggregate the capabilities of the machines in the tail of this distribution.

A virtual machine that aggregates the last 10 in Top500 would rank 32nd in ’95 but 14th in ‘03

Grid and P2P computing are, in part, results of this trend:Grids focus: infrastructure enabling controlled, secure resource sharing (for a relatively small number of resources) P2P focus: scale, deployability using integrated stacks.

Challenge: design services that offer the best of both worldscomplex, secure services, that deliver controlled QoS, are scalable and can be easily deployed.

Parameter 'k' evolution .

-0.84-0.82-0.80-0.78-0.76-0.74-0.72-0.70-0.68

1995

1996

1997

1998

1999

2000

2001

2002

2003

Roadmap

P2PImpactApplicationsMechanisms

PlanetLabWhy is this interesting for you?Compare resource management in PlanetLab and Globus

Wrap-Up

P2P Definition(s)

Def 1: “A class of applications that take advantage of resources (e.g., storage, cycles, content) available at the edge of the Internet.” (‘00)

Edges often turned off, without permanent IP addresses, etc.

Def 2: “A class of decentralized, self-organizing distributed systems, in which all or most communication is symmetric.” (IPTPS’02)

Lots of other definitions that fit in between

P2P impact today (1)

Widespread adoptionKaZaA – 360 million downloads (1.3M/week) one of the most popular applications ever!

leading to (almost) zero-cost content distribution:… is forcing companies to change their business models… might impact copyright laws

389,678Gnutella267,251MP2P

440,289Warez803,420iMesh

1,261,568Overnet1,987,097eDonkey2,460,120FastTrack

Sources: www.slyck.com, www.kazaa.com, July ‘04

0%

20%

40%

60%

80%

100%

Feb.'02 Aug.'02 Feb.'03 Aug.'03 Feb. '04 July'04

Other

Data transfers

Unidentified

File sharing

P2P impact today (2)

P2P – file-sharing - generated traffic may be the single largest contributor to Internet traffic todayDriving adoption of consumer broadband

Internet2 traffic statistics

Source: www.internet2.edu, July ‘04

P2P impact today (3)

A huge pool of underutilized resources lays around,users are willing to donate these resources• Seti@Home $1.5M / year in additional power consumed

which can be put to work efficiently (at least for some types of applications)

51.4 TFLOPSFloating point operations1.3K years1.3M yearsTotal CPU Time

23,3654,236,090Users764M

Total

1.13MResults received

Last 24 hours

Source: Seti@Home website, Oct. 2003

Roadmap

P2PImpactApplicationsMechanisms

PlanetLabWhy is this interesting for you?Compare resource management in PlanetLab and Globus

Wrap-Up

Applications: Number crunching

Examples: Seti@Home, UnitedDevices, DistributedSciencemany othersApproach suitable for a particular class of problems.

Massive parallelismLow bandwidth/computation ratioError tolerance, independence from solving a particular task

Problems:Centralized. How to extend the model to problems that are not massively parallel?

Relevant to Grid space: Ability to operate in an environment with limited trust and dynamic resources

Applications: File sharing

The ‘killer’ application to dateToo many to list them all: Napster, FastTrack (KaZaA, iMesh), Gnutella (LimeWire, BearShare), Overnet

Relevant to Grid space:Decentralized controlBuilding a (relatively) reliable, data-delivery service using a large, heterogeneous set of unreliable components.Chunking, erasure codes

0.8 TB/day on averageBytes transferred

~300,000Number of unique files≥ 10,000Number of local users

230,000/dayNumber of download sessions

Source: Israeli ISP, data collection 1/15-2/13/2003

FastTrack (Kazaa) load at a small ISP

Applications: Content Streaming

Streaming: the user ‘plays’ the data as as it arrives

source

Oh, I am exhausted!

Client/server approachP2P approach

Possible solution:The first few users get the stream from the serverNew users get the stream from the server or from users who are already receiving the stream

Relevant to Grid space: offload part of the server load to consumers to improve scalability

Applications: Performance benchmarking

Problem:Evaluate the performance of your service (Grid service, HTTP server) form end-user perspective

Multiple views on your site performance

Relevance to Grid space: Grid clients are heterogeneous, geographically dispersed Benchmark services for this set of consumers

End-to-end Performance Benchmarking

Back-end Infrastructure

Firewall

NetworkLandscape

Backbone

ISP

Consumer User

3rd partycontent

Web server

App server

Backbone

RegionalNetwork

Enterprise Provider

ISP

MajorProvider

Local ISP

T1

CorporateUserCorporate

Network

ComponentTesting

DatacenterMonitoring

End-to-endWeb PerformanceTesting

Database

Slide source: www.porivo.comSlide source: www.porivo.com

Many other P2P applications …

Backup storage (HiveNet, OceanStore)Collaborative environments (Groove Networks)Web serving communities (uServ)Instant messaging (Yahoo, AOL)Anonymous emailCensorship-resistant publishing systems (Ethernity, Freenet)Spam filtering

Mechanisms

To obtain a resilient system:integrate multiple components with uncorrelated failures, and use data and service replication.

To improve delivered QoS:move service delivery closer to consumer, integrate multiple providers with uncorrelated demand curves (reduces over-provisioning for peak loads)

To generate meaningful statistics, to detect anomalies:provide views from multiple vantage points

To improve scalability:Use decentralized (local) control, unmediated interactions

Roadmap

P2PImpactApplicationsMechanisms

PlanetLabWhy is this interesting for you?Compare assumptions and resource management mechanisms in PlanetLab and Globus

Wrap-Up

PlanetLab

Testbed to experiment with your networked applications.

>400 nodes, >150 sites, PlanteLab consortium: 80+ universities, Intel, HP

View presented to users: a distributed set of VMsAllocation unit: a slice = a set of virtual machines (VM), one VM at each node.

452 nodes162 sites450 research projects

VMM VMM VMM VMM

Slic

e K

OS

Slic

e K

OS

Slic

e K

OS

Slic

e K

OS

PlanetLab usage examples

Stress-test your Grid services (Globus RLS)GSLab: a playground to experiment with grid-services ‘Better-than-Internet’ services:

Resilient Overlays Multipath TCP (mTCP)Multicast Overlays

VMM VMM VMM VMMOS OS OS OS

Use

r acc

ount

Use

r acc

ount

Use

r acc

ount

Use

r acc

ount

Why should you find PlanetLab interesting?

1. Open, large-scale testbed for your P2P applications or Grid services

2. Solves a similar problem to Grids/Globus: building virtual organizations (or resource federations)

Grids: testbeds (deployments of hardware and software) to solve computational problems.

PlanetLab: testbed to play with CS applications

Main problem for both: enable resource sharing across multiple administrative domains

Roadmap

Summarize starting assumptions on: user communities, applications, resources and attempt to explain differencesLook at mechanisms to build VO.

Assumptions: User communities

PlanetLab: users are CS scientists that experiment with and deploy infrastructure services. Globus: users from a more diverse pool of scientists that are interested to run efficiently their (end-user) applications.

Implication: functionality offered

OS OS

User applications User applications

GlobusPL Services

PlanetLab.

Assumptions: Application characteristics

Different view on geographical resource distribution:

PlanetLab services: «distribution is a goal»leverage multiple vantage points for network measurements, or to exploit uncorrelated failures in large sets of components

Grid applications: «distribution is a fact of life»resource distribution: a result of how the VO was assembled (due to administrative constraints).

Implication: mechanism design for resource allocation

Assumptions: Resources

PlanetLab mission as testbed for a new class of networked services allows for little HW/SW heterogeneity.Globus supports a large set architectures; sites with multiple security requirements

Implications: complexity, development speed

Assumptions: Resource ownership

Goal: individual sitesretain control over their resourcesPlanetLab limits the autonomy of individual sites in a number of ways:

VO admins: Root access, Remote power buttonSites: Limited choice of OS, security infrastructure

Globus imposes fewer limits on site autonomyRequires fewer privileges (also can run in user space, )

PlanetLab emphasizes global coordination over local autonomy to a greater degree than Globus

Implications: ease to manage and evolve the testbed

PlanetLab

Globus

Individual site autonomy

Con

trol a

t VO

leve

l

Building Virtual Organizations

Individual node/site functionalityMechanisms at the aggregate level

Security infrastructureDelegation mechanisms

Resource allocation and schedulingResource discovery, monitoring, and selection.

Delegation mechanisms:Identity delegation

GlobusImplementation based on delegated X.509 proxies

PlanetLabNone

Broker/scheduler usage scenario:User A sends a job to a broker service which, in turn, submits it to a resource. The resource manager makes authorization decisions based on the identity that originated the job (A).

Broker

X.509/SSL

Delegated identity

Delegation mechanisms: Delegating rights to use resources

PlanetLabIndividual nodes managers hand out capabilities: akin to time-limited reservationsCapabilities can be traded Extra layer to: provide secure transfer, prevent double spending, offer external representation

GGF/GlobusWS-Agreements protocols:

To represent ‘contracts’ between providers and consumers.Local enforcement mechanism is not specified

The two efforts are complementary!

Broker/scheduler usage scenario:

User A acquires capabilities from various brokers then submits the job.

Delegated usage rights

BrokerResource

usage rights

Job descriptions +Usage rights acquired

Global resource allocation and scheduling

PlanetLabGlobus

Users

Application Managers

Brokers / Agents

Node Managers

Nodes(Resources)

• Resource usage delegation• Sends capabilities (leases)

• Identity delegation• Sends job descriptions

VO

ComputeCenter

VO(policies)

VOUsers

Services (runningon user’s behalf)

Access rights

Access

CAS or VOMSissuing SAMLor X.509 ACs

SSL/WS-Securitywith ProxyCertificates

Multiple VOs

Wrap-up

Scale & volatility

Functionality &infrastructure

Grids

P2P

Convergence: Large, Dynamic, Self-Configuring Grids

Large scaleIntermittent resource participationLocal control, Self-organizationWeaker trust assumptionsInfrastructures to support diverse

applicationsDiversity in shared resources

Fin

Links to papers, tech-reports, slides:Distributed Systems Group @ UChicago

http://dsl.cs.uchicago.edu

Thank you.

User

Userprocess #1

Proxy

Authenticate & create proxy credential

GSI(Grid

Security Infrastruc-

ture)

Gatekeeper(factory)

Reliable remote

invocation

GRAM(Grid Resource Allocation & Management)

Reporter(registry +discovery)

Userprocess #2Proxy #2

Create process Register

GT2 in One SlideGrid protocols (GSI, GRAM, …) enable resource sharing within virtual orgs; toolkit provides reference implementation ( = Globus Toolkit services)

Protocols (and APIs) enable other tools and services for membership, discovery, data mgmt, workflow, …

Other service(e.g. GridFTP)

Other GSI-authenticated remote service

requests

GIIS: GridInformationIndex Server(discovery)

MDS-2(Monitor./Discov. Svc.)

Soft stateregistration;

enquiry